🚀 go-pugleaf

This post is on behalf of a friend (Diana Fisher, New South Wales
National Parks and Wildlife service, Australia).

Q. I am using discriminant function analysis to find out which
variables are best able to discriminate between 2 animal groups,
using Systat.

My problem is that not enough information is available on each variable
describing the animals to give a normal distribution; most are in the 
form:

    Dependent variable = 0 (Not having the characteristic)

     or                = 1 (Having the characteristic)

although some have more classes, e.g. body weight which is in 6 classes.

Does this matter? I have a suspicion that body weight is being chosen by
the analysis as one of the best discriminators because it is in 6, rather
than 2, classes, unlike the other variables. Could this be right? Does
internal rescaling prevent this problem? Cross-tabulations do not show
a large difference between the weights of the 2 animal groups, and
changing body weight to 2 classes (Heavy vs. light) reduces the importance
of body weight as a factor.

Thanks in advance,

Diana.

Please e-mail responses to me. I will post a summary of responses to
sci.math.stat. Thanks.

Simon.
-- 
Simon Blomberg                                   simonb@extro.ucc.su.OZ.AU
School of Biol. Sciences, A08
University of Sydney           We demand the right to stop to look even at
NSW   2006  Australia          lizards, who are nobody's property. -JBS Haldane